UDI — Unified Developer Interface

A browser-based developer workbench that puts a multi-provider LLM chat interface, a sandboxed terminal, a file manager, and a configurable multi-agent pipeline all in one window.

---

Overview

UDI is a Flask application with a dark-themed single-page frontend. You type goals or questions into the Interface tab and an LLM responds; you switch to the Pipeline tab to run multi-agent workflows that can write files, lint code, run tests, and execute shell commands—all sandboxed to a local project/ directory.

---

Features

• Multi-provider LLM chat — supports Google Gemini, Anthropic Claude, and OpenAI GPT models, configurable via an environment variable.
• Multi-agent pipeline — define sequences of executor agents and an optional critic agent that iteratively refines output until the critic passes or a maximum iteration count is reached.
• Sandboxed code execution — agents can write files ([CODE]), run shell commands ([RUN]), lint ([LINT]), test ([TEST]), read files back ([READ]), and perform string replacements ([STR_RPL]), all confined to project/.
• Interactive [QUERY] pauses — a running pipeline can pause and ask the user a question, then resume with the answer incorporated.
• File API integration — files can be uploaded to the Gemini File API and referenced in chat.
• Retry logic — every LLM call retries up to 5 times with exponential backoff; Gemini calls run in a thread with a 45-second hard timeout.
• Live log panel — warnings and per-iteration agent logs stream to the UI.

---

Requirements

• Python 3.9+
• API key(s) for at least one provider (Gemini, Anthropic, or OpenAI)

Python dependencies

flask
flask-cors
google-genai
anthropic
openai

Install with:

pip install flask flask-cors google-genai anthropic openai

---

Configuration

All configuration is done via environment variables.

Variable	Description	Default
NARRATOR_MODEL	The model string to use for chat and pipeline agents	gemini-3.1-pro-preview
GEMINI_API_KEY	Google Gemini API key	(none)
ANTHROPIC_API_KEY	Anthropic API key	(none)
OPENAI_API_KEY	OpenAI API key	(none)

The provider is inferred automatically from the model name:

• Model names containing gemini → Gemini
• Model names containing claude → Anthropic
• Model names containing gpt → OpenAI

Example — use Claude:

export ANTHROPIC_API_KEY=sk-ant-...
export NARRATOR_MODEL=claude-sonnet-4-6

Example — use GPT:

export OPENAI_API_KEY=sk-...
export NARRATOR_MODEL=gpt-4o

---

Running

python app.py

The server starts on http://127.0.0.1:5000. Open that URL in a browser.

---

Project Structure

.
├── app.py          # Flask app, LLM handlers, /api/chat and /api/upload routes
├── pipeline.py     # Multi-agent pipeline engine, sandboxed execution, pipeline routes
├── templates/
│   └── index.html  # Single-page frontend (IBM Plex Mono, dark theme)
└── project/        # Sandboxed working directory for all agent file I/O (auto-created)

---

API Reference

Chat

POST /api/chat

Send a message to the configured LLM.

{
  "system_prompt": "You are a helpful assistant.",
  "context": "Optional background state.",
  "player_action": "The user's message.",
  "file_uri": "(optional) Gemini File API URI",
  "file_mime": "(optional) MIME type for the file"
}

Returns { "text": "...", "warnings": [] }.

---

POST /api/upload

Upload a file to the Gemini File API (multipart/form-data, field name file).

Returns { "uri": "...", "mime_type": "...", "name": "..." }.

---

GET /api/config

Returns the active model name and provider: { "narrator_model": "...", "narrator_provider": "..." }.

---

Pipeline

POST /api/pipeline/start

Start a pipeline job.

{
  "goal": "Build a web scraper for Hacker News.",
  "agents": [
    {
      "name": "Coder",
      "prompt": "Write the code to accomplish the goal.",
      "is_critic": false,
      "model": "gemini-3-flash-preview",
      "provider": "gemini",
      "temperature": 0.3
    },
    {
      "name": "Critic",
      "prompt": "Review the output and return JSON: { \"pass\": bool, \"feedback\": str }",
      "is_critic": true,
      "model": "gemini-3-flash-preview",
      "provider": "gemini"
    }
  ],
  "max_iterations": 3,
  "udi_context": "(optional) snapshot of the current UDI workspace state"
}

Returns { "job_id": "a1b2c3d4" }.

---

GET /api/pipeline/status/<job_id>

Poll for job state. Returns the full job object including status, logs, outputs, tags, run_outputs, feedback, and any pending query.

Status values: queued → running → completed / error / cancelled / paused

---

POST /api/pipeline/answer/<job_id>

Resume a paused job by answering a [QUERY].

{ "answer": "Use PostgreSQL." }

---

POST /api/pipeline/cancel/<job_id>

Cancel a running job.

---

GET /api/pipeline/jobs

List all jobs (summary).

---

File & Terminal

Method	Endpoint	Description
GET	/api/files	List all files in project/
POST	/api/files/read	Read a file: { "path": "..." }
POST	/api/files/write	Write a file: { "path": "...", "content": "..." }
POST	/api/files/delete	Delete a file: { "path": "..." }
POST	/api/run	Run a shell command: { "cmd": "..." }
POST	/api/lint	Lint a file: { "args": "filename" } or `"filename
POST	/api/test	Run tests: { "args": "filename" } or `"filename
POST	/api/read	Read a project file: { "filename": "..." }
POST	/api/str_replace	String replace: `{ "args": "filename

---

Pipeline Control Tags

Executor agents communicate intent through structured tags in their output. The pipeline engine parses and executes these automatically.

Tag	Syntax	Effect
[CODE]	[CODE: filename] ```lang code ``` [/CODE]	Writes code to project/filename
[RUN]	[RUN: command]	Executes a shell command in project/
[LINT]	[LINT: filename] or [LINT: filename | tool]	Lints a file (flake8 or eslint)
[TEST]	[TEST: filename] or [TEST: filename | type]	Runs pytest or jest
[READ]	[READ: filename]	Reads a file and injects its content into the next agent's context
[STR_RPL]	[STR_RPL: filename | old | new]	Performs a unique string replacement in a file
[QUERY]	[QUERY: question]	Pauses the pipeline and asks the user a question
[CONTEXT]	[CONTEXT: note]	Logs a reasoning annotation to the pipeline log

---

Security

All file operations and shell commands are sandboxed to the project/ directory. The following commands are always blocked regardless of context:

sudo, su, chmod, chown, dd, mkfs, fdisk, kill, pkill, reboot, shutdown, halt, poweroff, curl, wget, nc, ncat, netcat

Path traversal (..) and absolute paths outside project/ are also rejected.